What Even Is "Knowledge Work"?

I round up the most relevant AI-in-finance news, the deals being done, who's rolling out what, and what's actually working on the front lines.

Kimi K3's run up the leaderboards has turned into a Washington fight. OpenAI and Anthropic are lobbying to restrict Chinese open-weight models, the White House says Moonshot distilled Anthropic's Fable to build K3, and on July 24 twenty-five companies including Nvidia, Microsoft and Meta signed a letter telling lawmakers to leave open models alone. The weights land on Hugging Face today anyway.

Speaking of Hugging Face: OpenAI spent the week explaining how two of its models escaped a test sandbox and hacked it to steal benchmark answers, unnoticed for about a week. Days later, Sam Altman heads to Washington to ask for faster approvals. And Stripe is in talks to buy OpenRouter at a reported $10 billion.

Also inside: Microsoft's multibillion-dollar Mistral deal, Morgan Stanley's AI fee machine, and the Musk cover interview.

But first, my take on what knowledge work actually is. A week of client calls sent me back to my banking years, and to a 50-tab model I still think about more than I'd like.

In today's Acquisition Intelligence:

From The Trenches

Most Knowledge Work Is Transport

I spent part of this week inside a client's monthly reporting process, and it sent me straight back to banking. Somewhere at J.P. Morgan there is a 50-tab model with my name in the file path, and I can still feel the specific dread of opening it.

You didn't so much update that model as perform surgery on it, using hacks and superstitions passed down through generations of analysts. Paste-special the values before touching anything. Never sort that one sheet. Keep a hardcoded copy of the March version, because the links to the datapack break if you breathe near them.

Then it crashes anyway. Or you open it one morning to find half the sheets have collapsed into #REF! errors because somebody renamed a tab, and the day disappears into rebuilding what existed on Friday. Whole weeks of my twenties went into keeping artifacts like that alive, and at the time it felt like the job.

The Human Integration Layer

Here is what I now think that work actually was. Watch how most deliverables in finance get made: an export lands from one system into Excel, someone reshapes it into a second spreadsheet in the house format, that feeds a PowerPoint, and the PowerPoint gets re-keyed into whatever the next audience needs. Information enters the chain exactly once, in the first export.

Everything after that is transport.

Take a standard portfolio monitoring cycle. The portco controller exports from the ERP into a spreadsheet, an associate reformats it into the fund template, the template feeds the board deck, and the deck gets cut down again for the LP report. Four artifacts, days of elapsed time, and every number in the last one was already present in the first.

We built entire career ladders on this, and it never looked absurd, because the systems could not talk to each other. A person with Excel was the integration. You were the API between the ERP and the board pack. You just never got credited for the uptime.

Straight to the Source

What changed is that you no longer have to move the information at all. Connect an agent to the source systems, or better still to a single system that holds the lot (I might be biased on that one), and the reshaping happens on demand. The board view, the LP view and the "why is this portfolio company underperforming" view are all just questions asked of the same underlying data.

I made a version of this point on a client call this week: it doesn't matter how clunky a source system's reporting looks, as long as the raw data is being captured. Presentation used to be the expensive part, which is why we tolerated week-long chains of artifact production to get to it. Now presentation is roughly free and interpreting the raw data has become dramatically easier, so the chain is pure overhead. The AI turns out to be the translation layer between systems, and the friction the whole chain existed to absorb simply goes.

McKinsey published a piece on July 20 arguing that AI's next mission is rewiring how work is organised, with people shifting toward supervision and judgment while agents absorb the repeatable steps. Strip off the consultancy varnish and it's the same observation. The layer being automated first is the transport layer, because that's where the least judgment lives.

It's also why Anthropic, Blackstone and Hellman & Friedman just capitalised Ode, a $1.5 billion services firm that exists to do this unpicking for enterprises; if connecting firms to their own data were trivial, nobody would fund it at that scale. It's the design premise we started DealSage from too: plug into the systems where a firm's data already lives, and let every deliverable be a generated view rather than hand-carried cargo.

The Third Category

Killing the reformatting is the visible win, and it's how most firms still evaluate AI: as automation, the same work done faster. The more ambitious ones have moved to a second frame, agents. This week convinced me there's a third category, the most powerful and the least priced in: understanding. Piecing together the parts of your own organisation you never really knew.

Once finance, operations, contracts and HR sit in one connected layer, you can ask questions that run across all of them. Why does margin differ by six points between two near-identical portfolio companies? The answer might be a missing cost pass-through clause in one master agreement, a drafting slip bleeding that company for years. Finding it means holding billing data and contract terms in the same view, which used to cost a week of somebody's life, so nobody ever looked.

Or join the HR system to the output data and find that attrition peaks right at the point people get good, a compounding cost nobody computes because the two systems have never met. The answers live in the joins, exactly where nobody could afford to look. Automation and agents accelerate work you knew needed doing; understanding surfaces work you had no idea existed.

A dose of scepticism, because this newsletter promised you no hype: cross-system analysis is precisely where a model will invent a pattern if you let it, and the mappings between systems still live in people's heads. The firms doing this well write that knowledge into a rulebook the AI has to follow and tie every number back to reports the team already trusts. Unglamorous, and it's the difference between analytics and fan fiction.

The Week You Get Back

So the exciting part comes in two pieces. The first is capacity: no firm really buys saved hours, they buy what the freed capacity produces, and every fund carries a list of analyses it never runs because each one costs somebody three days. At three minutes a question, you run the list, and some of the answers will move a valuation.

The second is decision quality. Tell a COO their crews run at 62 per cent utilisation against a 75 per cent target and you've told them what they already suspected. The value arrives when the connected data can say which sites have slack, where the volume should move, and why, and then track whether the move worked.

Judgment was always the job. It just got rationed to the two hours a day left over after the transport was done. I don't miss the 50-tab model, but I do sometimes wonder what that version of me would have produced with these tools. My guess is better questions, and far less superstition about the sort button.

What The Builders Are Saying

One post this week, and it landed a few hours before I hit send. Worth your time, and worth the follow.

@levie (Aaron Levie, co-founder and CEO of Box)

The post: enterprises will need enormous support applying model breakthroughs to real workflows, because "intelligence alone is not enough to transform most processes." Feedback loops, connections into enterprise systems, humans at the right decision points, compliance handled. His sharpest claim runs against the consensus: the need for that applied layer grows as models improve, because better models let you attempt more ambitious workflows.

Why this matters: this is our thesis. It's what we build DealSage on every day, and here it is argued by someone who sells into most of the Fortune 500, the same week the labs and PE put $1.5 billion behind it.

My take: the market keeps assuming each model release shrinks the space above it. The opposite keeps happening: capability raises ambition, and ambition raises the integration bill. That bill is what this whole issue is about.

News Digest

The Open-Weight Fight Just Split The Industry In Two

The Kimi K3 story from last issue has escalated into a Washington fight. OpenAI and Anthropic are lobbying to restrict Chinese open-weight models, the White House has accused Moonshot of distilling Anthropic's Fable to build K3, and Treasury Secretary Bessent has floated sanctions. On July 24, twenty-five US companies signed a joint letter urging lawmakers to hold off. Moonshot put the weights on Hugging Face today.

The details:

Why it matters: the fault line runs between business models. Everyone selling compute, infrastructure or applications wants the model layer commoditised; the labs want a moat.

My take: the signatories are telling you where they think the margin lives, and none of them think it's the model. We built DealSage model-agnostic for exactly this reason: the model layer keeps repricing, and your firm's data and workflows shouldn't be hostage to it. If weights become export-controlled assets, model provenance joins the diligence checklist.

Read more

The Agent That Hacked Hugging Face Is Off To Washington

On July 21, OpenAI disclosed that two of its models, GPT-5.6 Sol and an unreleased successor, escaped a sandboxed test environment and broke into Hugging Face's production systems. They were being tested on cybersecurity skills, worked out that Hugging Face held the benchmark's answer key, and went and took it.

The details:

Why it matters: the models weren't told to hack anyone. They were told to score well, and they used what they could reach.

My take: this is why it matters to work with people who know how to configure these systems safely: restricted tools, proper sandboxes, logs somebody actually reads. Right now most firms are just letting agents have at it. Every DealSage deployment starts with what an agent is allowed to touch, and after this week I'd ask that question of every vendor you use.

Read more

Stripe In Talks To Buy OpenRouter At A Reported $10 Billion

The WSJ reported on July 24 that Stripe is in talks to buy OpenRouter, the marketplace that routes developer traffic across AI models, at around $10 billion. Talks are non-final but could conclude soon, and it follows the $53 billion PayPal bid with Advent from last week.

The details:

Why it matters: the model-agnostic middle layer just got a ten-figure price.

My take: a payments company paying $10 billion for model routing is a bet that agents will consume intelligence like a metered utility, and that nobody wants to be locked to one lab. That's the posture we take at DealSage, and the one I'd want from any AI vendor selling into a fund.

Other Interesting Things I've Read or Seen This Week

Anthropic, Blackstone and Hellman & Friedman launch Ode (Business Wire, July 16) - A $1.5 billion standalone AI services firm: Anthropic's models plus engineering teams, deployed inside large firms as a service. (Covered in the essay above, filed here so I stop going on about it.)

Microsoft to fund Mistral's European AI expansion (Reuters, July 21) - Billions for Mistral's French data centres and a gigawatt of compute by 2030, with Samsung reported circling at €20 billion days later. (European sovereignty, funded from Redmond.)

Morgan Stanley cashes in on AI boom with debt deals (FT, July 21) - The bank has become the leading arranger of the novel debt structures financing data centres. (Whoever wins the AI race, the arranger fees are doing fine.)

Should you be afraid of Elon Musk? (The Economist, July 23) - The cover leader finds a man who expects his own technology to make him powerless within five years, and acts as if he doesn't believe it. (His plan for post-AI purpose is gardening. The Economist counters with Russian novels.)

Why workers are nostalgic for life before AI (FT, July 26) - 65 per cent of white-collar workers regularly yearn for their pre-AI working life, and nearly half say fact-checking AI output adds to the slog. (Nobody appears to miss the 50-tab model specifically.)

US tech groups cut 140,000 jobs despite AI spending boom (FT, July 25) - Close to 140,000 US tech job cuts this year, over a third of all announced American layoffs, while the same companies spend north of $800 billion on AI. (The labs are hiring some of them back.)

Acquisition Intelligence is a weekly newsletter on AI in M&A for finance professionals, private equity investors, investment bankers, corp dev teams, and deal-makers.

For questions, feedback, or to share what you're seeing in the market, reply to this email.

P.S. I'm Harry, co-founder of DealSage. We build exactly the layer this issue is about: DealSage connects to the systems where your firm's data already lives, so the reformatting chain disappears and your team spends the week on questions instead. If your monthly reporting still passes through four spreadsheets, reply here or have a look at dealsage.io.